Do paraphrased constraints actually survive, or only exact wording?

Mostly. On probe words Jev holds about 0.87 on held-out paraphrases versus 0.86-0.88 canonical, while the keyword checker drops 16-24 points. Expect the same judgment pipeline, not identical wording.

What is the catch with near-miss spatial constraints?

That is the documented weak spot: 38-53% false positives on near-miss negatives such as left edge versus top edge. The confidence gate and max-over-steps aggregation exist to contain it, not remove it.

Can I afford this inside a 300-candidate CEM loop?

About 2.5s per 1,000 judgments, around 5.5 calls and 31k input tokens per episode with step-fact memoing across 300 candidates, iterations, replans, and runs. Soft constraints judge only the 16 lowest-cost candidates.

Why was the first randomized study null for every judge?

In random starts 53% of spatial episodes place goal or start inside the forbidden cell, so success and compliance conflict even for the oracle. The positive results come from starts where compliance is possible.

How is a violation turned into a penalty?

Three families, all folded in Python. never asks as a Noul whether the block or agent violates the constraint at step t and aggregates max over steps. always asks whether the facts are consistent with the constraint and aggregates 1 minus the minimum. soft asks as a five-level Score how well the candidate respects it overall and returns 1 minus the expected level over four. Cost is the standardised latent distance plus a weighted sum of those penalties, and at lambda zero it is bit-identical to plain LeWM.

Why does it beat a local LLM on the same job?

Partly because the comparison is not close. A local Qwen2.5-7B was run on 16 items and scored 0.73 accuracy on the one call in three that returned valid JSON, at 566 seconds per 1,000 judgments. Jev scored 0.86 to 0.88 with AUROC 0.94 at 2.5 to 2.6 seconds per 1,000. The oracle-on-probes upper bound sits at 0.97 to 0.99, so the headroom above Jev is real but modest.

If the judge is not the problem, what is?

The world model. With a perfect judge, the chosen plan is imagined to enter the centre cell in 43% of spatial episodes and actually does in 92%, so the remaining gap is drift in the imagined rollout rather than judging. Probe word accuracy decays from 0.95 to 0.88 over five imagined steps, which is why six of nine pre-registered hypotheses were met while the oracle's own 20-point target was not.

CommunityJev Tools & CLIs

Lejudge Jev Jepa

English-programmable cost module for JEPA PushT planning: linear probes verbalize imagined latents and Jev judges each constraint step, folded into CEM cost with memoised step facts.

See highlights

Open github.com

Stars: 10
Language: Python
License: MIT
Published by: Community
Author: AbdelStark
Last verified: Oct 2026

AbdelStark/lejudge-jev-jepa

User repository on GitHub

View on GitHub

Natural-language constraints for JEPA world-model planning, judged by a decision model instead of an LLM.

Forks: 5
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Natural-language constraints for JEPA world-model planning, judged by a decision model instead of an LLM. Imagine → describe → judge → decide. LeWorldModel (LeWM) plans by minimising one number: the latent distance between an imagined final state and a goal image. That cost cannot say "but not through there" or "don't touch it yet". LeJudge adds a second term that anyone can write in plain English: Imagine. LeWM's predictor rolls out 300 candidate action sequences in latent space (CEM, 30 iterations).

Quoted from the project's README. Read the full README.

Tagscommunity awesomejev catalog-import

Works forJev for Classification Jev for Evaluation

Used inJev for Software & DevTools

Languages: Python

Highlights

LeWM imagines 300 sequences; probes map each latent to grid cell, edge, angle bin, contact, and speed words.
Jev answers one typed yes-no per constraint and step; Python folds never as max, always as 1-minus-min, soft as expected level.
Owner text and generated facts stay in separate fields, so an injected constraint is judged as one nothing violates.
Identical-code oracle on the same probe words attributes each miss to judge, cost mechanism, or world model.

Watch out

Study code and cached judgments reproduce the paper offline, but compliant starts are curated and world-model drift remains the larger error term. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jevclip

Video tool where Jev scores subtitle segments and code assembles a short highlights cut plus a de-watered full cut with timestamped summaries.

Jev Tools & CLIs#community#jevbest#catalog-import

Community

cclank/jevclipPython

jev-dimabsa

Research code that tackles dimensional sentiment scoring with Jev typed Score, Choice, and Noul calls plus small corpus statistics, without generation, fine-tuning, or GPU.

Jev Tools & CLIs#community#awesomejev#catalog-import

Community

ZhangYiqun018/jev-dimabsaPython

antivirus

Windows-first scanner and EDR triage tool that extracts structural features from binaries, scripts, documents and URLs, then has Jev return a severity score and a verdict for each target.

Jev Tools & CLIs#community#awesomejev#catalog-import

Community

newuser7171/antivirusPython

jev-trade-cc

A point-in-time backtest of a William O'Neil momentum strategy where jev picks the entries and decides the exits, paired against the same rules with no model — and honest enough to report the model as indistinguishable from them.

Jev Tools & CLIs#community#awesomejev#catalog-import

Community

michaelpersonal/jev-trade-ccPython

typesafe-comment

A Python linter that scores code comments with System One on five heuristics — usefulness, readability, accuracy, redundancy, coverage — and fails the pipeline when a comment falls under a threshold.

Jev Tools & CLIs#community#jevbest#catalog-import

Community

Hexdigest123/typesafe-commentPython

jev-seo

547

Live SEO audit that crawls a site from one homepage URL, checks 52 rules tied to Google Search Central, and asks Jev typed questions about every page to rank fixes.

Jev Tools & CLIs#community#jevbest#catalog-import

Community

AgriciDaniel/jev-seoPython

From the Jev guides

Original write-ups on what this entry is for.

Jev on the Gateways: Vercel, Workers AI, and OpenRouter

Oct 12, 2026

The request shape barely changes across surfaces; the surrounding contract does. When to call the API directly and when a gateway, a Worker binding, or a router entry earns its extra hop.

Jev Tools & CLIs#gateway#vercel#cloudflare

Read article

Exporting Decision Models: GGUF, ONNX, and Device Runtimes

Oct 10, 2026

Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.

Jev Tools & CLIs#open-models#local#onnx

Read article

Records at Rest: Documents, Inboxes, and Risk Queues

Sep 28, 2026

Jev labels a record that already exists. Code moves, flags, or holds it. How the document, mail, and transfer tools keep the action out of the model.

Jev Tools & CLIs#classification#email#security

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai